---
title: 'RDTE-UNet: Hybrid Medical Segmentation'
url: https://www.emergentmind.com/topics/rdte-unet
type: topic
---

# RDTE-UNet: Hybrid Medical Segmentation

RDTE-UNet is a neural architecture designed for precise medical image segmentation, targeting enhanced boundary accuracy and preservation of fine anatomical details. It achieves this by integrating local convolutional modeling with global Transformer-based context awareness. The architecture introduces three principal innovations: the Adaptive Shape-aware Boundary Enhancement (ASBE) module for edge sharpening, the Horizontal-Vertical Detail Attention (HVDA) block for fine-grained directional feature modeling, and the Euler Feature Fusion (EulerFF) module for direction- and channel-sensitive skip connection fusion. Evaluated on Synapse multi-organ CT and BUSI breast ultrasound datasets, RDTE-UNet demonstrates advanced segmentation performance in both quantitative and qualitative terms [2511.01328].

## 1. Architectural Overview

RDTE-UNet adopts a U-shaped encoder–decoder topology, comprising five encoder and five decoder stages connected via classical skip-links. The encoder alternates between standard ResBlock stages (1–3) and "Details Transformer" stages (4–5). The architecture is organized as follows:

- **Front-end ASBE Module**: Replaces the initial convolution with an edge-aware block. ASBE utilizes Adaptive Rectangular Convolution for shape-dependent kernel learning and a boundary-difference operator for edge sharpening.
- **Hybrid Encoder Backbone**: Stages 1–3 are implemented as ResBlocks for strong local representation. Stages 4–5 transition to a Details Transformer block comprising HVDA—emphasizing horizontal and vertical features using directional “StairConv” mechanisms—followed by a channel MLP and Pre-LN residual structures for global modeling.
- **Decoder Path and Fusion**: At each upsampling stage, corresponding encoder features are fused with the decoder stream via the EulerFF module, allowing dynamic weighting of horizontal, vertical, and channel responses through a complex-valued Eulerian formulation.
- **Output Head**: A final $1\times1$ convolution maps integrated features to class logits for segmentation.

## 2. Module Formulations

The RDTE-UNet's three chief modules are mathematically defined as follows:

### 2.1 Adaptive Shape-Aware Boundary Enhancement (ASBE)

Given $x\in\mathbb{R}^{H\times W\times C}$,
\[
f_0 = \mathrm{Conv}_{1\times1}(x)
\]
Feature branches:
\[
f_{\mathrm{pool}} = \mathrm{AvgPool}_{s}(f_0), \qquad f_{\mathrm{ar}} = \mathrm{ARConv}(f_0)
\]
Boundary-difference enhancement:
\[
\Delta f = f_{\mathrm{pool}} - f_{\mathrm{ar}}, \qquad f_e = \sigma(\alpha\, \Delta f) \odot f_{\mathrm{ar}}
\]
Channels fused:
\[
f_{\mathrm{out}} = \mathrm{Conv}_{1\times1}\left(\left[f_e,\, f_0\right]\right)
\]

### 2.2 Horizontal-Vertical Detail Attention (HVDA)

For $x_\mathrm{in}\in\mathbb{R}^{h\times w\times d}$:
\[
x_{hd} = \mathrm{StairConv}_h(x_\mathrm{in}), \qquad x_{vd} = \mathrm{StairConv}_v(x_\mathrm{in})
\]
Concatenation and fusion:
\[
x_{\mathrm{cat}} = \mathrm{ReLU}\left(\mathrm{BN}(\mathrm{Conv}_{1\times1}([x_{hd}, x_{vd}]))\right) + x_{hd} + x_{vd}
\]
\[
x_{\mathrm{fuse}} = \mathrm{ReLU}\left(\mathrm{BN}(\mathrm{Conv}_{3\times3}(x_C)) + x_C\right) + x_{hd} + x_{vd}
\]
Self-attention module:
\[
Q = W_Q x_{\mathrm{fuse}},\ K = W_K x_{\mathrm{fuse}},\ V = W_V x_{\mathrm{fuse}}
\]
\[
B = \mathrm{softmax}\left(\frac{QK^T}{\sqrt{d_k}}\right),\quad \mathrm{HVDA}(x_{\mathrm{in}}) = BV
\]
StairConv operations use asymmetric, multi-scale padded convolutions for directional feature extraction.

### 2.3 Euler Feature Fusion (EulerFF)

Given features $x_s$ (skip) and $x_d$ (decoder), project both into a complex Eulerian domain:
\[
\mathcal{F}^{(h)} = A_h \cos(\theta_h) + jA_h \sin(\theta_h)
\]
\[
\mathcal{F}^{(v)} = A_v \cos(\theta_v) + jA_v \sin(\theta_v)
\]
Apply group convolutions separately to real and imaginary parts for horizontal and vertical branches, aggregate channel-wise, and fuse:
\[
\mathcal{F}_{\mathrm{out}} = \mathrm{Conv}_{1\times1}\left([x_s, x_d, \widetilde{\mathcal{T}}_h, \widetilde{\mathcal{T}}_v, \widetilde{\mathcal{T}}_c]\right)
\]

## 3. Network Assembly and Layerwise Pipeline

The layerwise pipeline is as follows:
```python
# Input: image I ∈ R^{H×W×C}
f0 = ASBE(I)                   # → R^{H×W×C'}
e1 = ResBlock(f0)              # H/2, W/2, 2C'
e2 = ResBlock(e1)              # H/4, W/4, 4C'
e3 = ResBlock(e2)              # H/8, W/8, 8C'
e4 = DetailsTransformer(e3)    # H/16, W/16, 16C'
e5 = DetailsTransformer(e4)    # H/32, W/32, 32C'

d5 = Deconv(e5)                # H/16, W/16, 16C'
d5 = EulerFF(skip=e4, dec=d5)
d4 = Deconv(d5)                # H/8, W/8, 8C'
d4 = EulerFF(skip=e3, dec=d4)
d3 = Deconv(d4)                # H/4, W/4, 4C'
d3 = EulerFF(skip=e2, dec=d3)
d2 = Deconv(d3)                # H/2, W/2, 2C'
d2 = EulerFF(skip=e1, dec=d2)
d1 = Deconv(d2)                # H, W, C'
d1 = EulerFF(skip=f0, dec=d1)

out = Conv_{1×1}(d1)           # H×W×N_classes
return out
```
This schema reflects the mixed convolutional-transformer design, edge-aware enhancement, and direction-sensitive post-processing prior to segmentation output.

## 4. Training Protocols and Hyperparameters

- **Loss Function**: Combination of Dice loss and cross-entropy
  \[
  \mathcal{L} = \lambda_{\mathrm{Dice}} (1 - \mathrm{Dice}(P,T)) + \lambda_{\mathrm{CE}}\,\mathrm{CE}(P,T)
  \]
  with $\lambda_{\mathrm{Dice}} = \lambda_{\mathrm{CE}} = 0.5$.
- **Optimizer**: Adam, parameters $(\beta_1=0.9,\, \beta_2=0.999)$, initial learning rate $1\times10^{-4}$, weight decay $1\times10^{-5}$.
- **Learning Rate Scheduling**: Cosine annealing with warm restarts (SGDR), training for 200 epochs.
- **Batch Sizes**: 8 (Synapse), 16 (BUSI).
- **Data Augmentation**: Random rotations (±15°), flips, intensity scaling.

## 5. Quantitative and Qualitative Results

On Synapse CT (multi-organ) and BUSI breast ultrasound benchmarks, RDTE-UNet achieves the following:

### Synapse Dataset (60/40 split)
| Method         | DSC (%) ↑ | HD95 (mm) ↓ |
|----------------|-----------|-------------|
| Trans-UNet     | 79.15     | 28.47       |
| Swin-UNet      | 81.03     | 19.54       |
| MT-UNet        | 80.72     | 22.48       |
| RWKV-UNet      | 85.62     | 14.83       |
| RDTE-UNet      | 86.63 ★   | 11.69 ★     |

### BUSI Dataset (70/30 split)
| Method         | DSC (%) ↑ | HD95 (mm) ↓ |
|----------------|-----------|-------------|
| Trans-UNet     | 60.42     | 32.78       |
| Swin-UNet      | 62.91     | 30.67       |
| MT-UNet        | 62.13     | 39.08       |
| RWKV-UNet      | 64.85     | 29.57       |
| RDTE-UNet      | 66.31 ★   | 27.73 ★     |

The superior boundary quality and structural consistency are substantiated by sharper edges and fewer false positives, particularly around complex morphologies such as the pancreas and fine-grained vasculature.

## 6. Context and Comparative Significance

RDTE-UNet advances the segmentation performance over previously reported methods by coupling explicit boundary enhancement (ASBE), refined directionality (HVDA), and mathematically principled fusion (EulerFF). The improvements are most pronounced in boundary-sensitive metrics (HD95) and Dice similarity coefficient (DSC). Its hybrid convolution-transformer backbone distinguishes it from pure CNN or typical hybrid architectures by its two-stream encoder and detail- and shape-aware auxiliary branches. This suggests the benefit of modeling interactions between local and long-range detail, especially when segmenting anatomically complex or ambiguous regions.

A plausible implication is that future segmentation methods in computational medicine may increasingly incorporate direction- and boundary-aware modules alongside global context mechanisms to simultaneously address fine structure delineation and region-level accuracy.

Source: https://www.emergentmind.com/topics/rdte-unet